Press Coverage and Political Accountability
James M. Snyder, Jr,∗and David Strömberg†
March, 2008
Abstract
In this paper we estimate the impact of press coverage on citizen knowledge, politicians’ actions, and policy. We find that a poor fit between newspaper markets and
political districts reduces press coverage of politics. We use variation in this fit due
to redistricting to identify the effects of reduced coverage. Exploring the links in the
causal chain of media effects — voter information, politicians’ actions and policy — we
find statistically significant and substantively important effects. Voters living in areas with less coverage of their U.S. House representative are less likely to recall their
representative’s name, and less able to describe and rate them. Congressmen who are
less covered by the local press work less for their constituencies: they are less likely to
stand witness before congressional hearings, to serve on constituency-oriented committees (perhaps), and to vote against the party line. Finally, this congressional behavior
affects policy. Federal spending is lower in areas where there is less press coverage of
the local members of congress.
∗
Department of Political Science and Department of Economics, Massachusetts Institute of Technology.
email: millett@mit.edu
†
Institute for International Economic Studies, Stocholm University. email: david.stromberg@iies.su.se
1
1
Introduction
It has long been recognized that having a press that actively covers politics is essential for
democratic governance. Informed voters are better able to hold elected officials accountable
for their policy decisions, and most people get their information via the media. In the words
of Thomas Jefferson:
“The functionaries of every government have propensities to command at
will the liberty and property of their constituents. There is no safe deposit for
these but with the people themselves, nor can they be safe with them without
information. Where the press is free, and every man able to read, all is safe.”
—Thomas Jefferson to Charles Yancey, 1816. ME 14:384
Frequently, however, an active press is missing from the political stage. At the national
level, this is true in many weak democracies and poor countries, often because of political
oppression or lack of an economic base to support large mass media. This is true also in
mature democracies with a free press, in particular for lower levels of government. One reason
is that local politics is boring to many people. Another reason it that there is often a poor fit
between media markets and political jurisdictions, making coverage of some jurisdictions too
costly. This problem may be increasing with increasing media market size, consolidations
in the media industry and a shift from print to electronic media. As the number of active
political units within each media market that require media attention increases, reporting,
by necessity, becomes more selective and superficial. To counter this trend, regulations such
as the Newspaper Preservation Act and the FCC “must carry” laws aim at ensuring local
content and programming. For similar reasons, many governments subsidize domestic media.
Despite these important and long standing concerns, empirical evidence of the effects
of active media coverage is scarce. One reason is that media coverage is endogenous to
most outcome variables of interest. For example, a number of studies have found better
governance in countries where the media actively covers politics (Ahrend, 2002, Brunetti
and Weder, 2003, Djankov and others, 2003). However, allowing active political coverage
is a governmental choice, and more corrupt governments have stronger incentives to silence
the press. So, correlations between governance and active political coverage may simply
reflect that these variables are jointly determined, as discussed by Besley and Prat (2006).
Similarly, while some studies find a correlation between media coverage and voter knowledge,
this correlation might simply reflect the fact that news coverage and knowledge are both
driven by citizens’ intrinsic interest in politics.
One way to more convincingly identify an effect of active media coverage on voters,
politicians, and policy, is to find a number of similar political jurisdictions where some factor
2
causes news coverage of politics to vary, but this factor does not directly affect our outcome
variables. The factor we propose is the match, or “congruence,” between newspaper markets
and U.S. congressional districts. The “economic geography” factors that determine media
markets are generally quite different from the “political geography” factors that determine
congressional district boundaries. Therefore, it therefore seems a priori reasonable that
the match between the two will be haphazard. Below we present some evidence to this
effect. In addition, panel features of our data allow us to estimate two specifications that are
particularly convincing (at least to us). One investigates the consequences of the changing
congruence between media markets and congressional district due to redistricting. The other
employs congressional-district-and-year fixed effects, and thus focuses on variation within
congressional districts.
The paper follows the chain of media impacts, link by link: congruence increases the
newspapers’ political coverage which, in turn, makes voters better informed, which increases
monitoring and induces politicians to work harder, which, finally, produces better policies
(for their constituencies). Figure 1 shows these links, together with simple bivariate graphs
of the main relationships of interest.1 The paper is devoted to investigating whether causal
relationships are underlying the patterns in those graphs.
We start by defining the match, or congruence, between newspaper markets and congressional districts in section 2. The measure we use is based on the share of a newspaper’s
readership that lives in a certain congressional district. In Figure 1(a), the left image shows
a perfect match between media markets and congressional districts. Every person in both
districts only buys newspapers that only sell in this district. The situation to the right shows
a worse match where each person reads a newspaper which only has half of the sales in their
district. Intuitively, newspaper coverage of a congressman should be increasing in this readership share. Since more than one newspaper sells in each district, we define congruence as
the circulation-weighted average of readership for all newspapers sold in a district.
We then explore whether newspaper coverage is increasing in the readership share in
section 3. Using online editions, we counted the number of articles mentioning House representatives in 161 newspapers which, on average, cover 385 districts in each congress from
1991-2002. We find that coverage of representatives is strongly increasing in district readership shares. This can clearly be seen from the graph in Figure 1(b), which plots the share
of articles mentioning a given representative in a given newspaper as a function of the share
of that newspaper’s readers that lives in the representative’s district.
Consequently, voters in non-congruent areas are less exposed to news about their rep1
Except for the first graph, each point in the graphs the mean value of the y-variable within each percentile of Congruence. So, for example, the first point in the “Witness appearances” graph shows the mean
Congruence and witness appearances in the lowest Congruence percentile.
3
resentatives. The newspapers sold in their area simply devote less coverage to their House
representative. Section 4 explores whether this affects voter information. After analyzing
survey responses from the American National Election Studies 1984 to 2004, we find that
voters from more congruent areas are better informed about their House representative.
More precisely, they are better able to correctly name at least one of the candidates in the
House election. They are also more willing to place their representative ideologicially, to rate
their representative on a feeling thermometer, and to mention things that they like or dislike
about their representative. The graph in Figure 1(c) plots the share that can correctly recall
a name against congruence. Since voter information is key to accountability, this relationship
is extensively explored. In an extension, we also study how news coverage affects two voting
outcomes, participation in elections and the tendency to support the incumbent.
Section 5 then studies effects on representatives’ behavior. Some representatives are
elected from districts that “fall through the cracks” of some or all local newspapers, and the
press is largely silent about them. Others are elected from highly congruent districts with
ample coverage and well-informed voters. Such representatives might therefore have a greater
incentive to work for their constituency, for example, by considering constituency (rather
than party) interests in voting, and by engaging in pork-barrel politics. One useful vehicle
in bringing pork to the district is to serve on “constituency” committees with jurisdiction
over spending programs, another to appear as a witness before congressional hearings. To
analyze these aspects of representative effort, we collected roll-call voting data for all House
members 1982-2002, and data on committee assignments and witness appearances for 19822004. We find that congressmen from highly congruent districts are more disciplined by
their constituencies: their voting conforms less to the party line, they are more likely to
stand witness before congressional hearings, and they are, perhaps, more likely to serve
on constituency-oriented committees and also less likely to serve on broad policy-oriented
committees. Figures1 (d),(e), and (f) show the corresponding bivariate graphs.
If the representatives’ efforts in Congress are successful, then we should observe more
federal funds flowing into highly congruent districts, and into more congruent counties within
districts. To test this, we assembled data on federal expenditure allocations across counties
from the Consolidated Federal Funds Report, 1983-2004. The total value of the expenditures
we study is about $2,700 per capita (in year 2000 dollar values), or about 10 percent of U.S.
GDP. We find that more congruent counties have received higher federal expenditures per
capita. Figure 1(g) shows the corresponding bivariate graph.
As mentioned, this paper is closely related to the literature on the political consequences
of having a media that actively covers politics. Moreover, our results in section 5 are related
to Cohen, Noll, and Zaller (n.d.), who study how the amount of information available to
voters affects congressional roll call voting behavior, using an approach which is similar to
ours. Consistent with our results, they find evidence that representatives in low-information
4
districts vote in a more purely partisan manner, while those in high-information districts
are more responsive to the underlying ideology of the district. Our work is also related to
several paper that use television media market definitions to explore the importance of mass
media on the incumbency advantage and campaign finance.2
While we study the effects of media content, our results are related to a number of studies
of the effect of media penetration and media use. For example, our results in section 4 are
closely related to a literature on the effects of mass media use on information.3 However,
while there are numerous studies regressing measures of political knowledge on self-reported
media exposure (plus other controls), none of them convincingly identifies a causal relationship. One problem is that people who are interested in politics consume more political news,
from a variety of sources, causing political knowledge and news consumption to be positively correlated. The results in section 6 are related to the findings that media penetration
has been found to influence government spending (Strömberg, 1999, 2004, and Besley and
Burgess, 2002). Our approach is different since we study the effects of news coverage, holding
media penetration constant.
Before proceeding, we must briefly discuss why we focus on newspapers and ignore radio
and television. The existing evidence indicates that local television stations devote much
less coverage to congress than local newspapers (Hess, 1991; Vinson, 2003). In one analysis,
we compare the impact of television market structure and newspaper market structure on
voter knowledge (section 4), and find that television appears to have no effect on voter
knowledge about their congressmen. Less is known about radio, and this will require further
investigation.
2
Congruence
The driving force behind all results in this paper is that the number of articles, qmd , that a
newspaper m writes about a House representative from district d, is strongly increasing in the
share of this newspaper’s readers that lives in district d, ReaderSharemd . We present strong
evidence of this in the empirical section. For simplicity, assume a proportional relationship,
so that
qmd = αReaderSharemd .
(1)
2
See Campbell, Alford and Henry (1984), Niemi, Powell and Bicknell (1986), Stewart and Reynolds
(1990), Levy and Squire (2000), Ansolabehere, Snowberg and Snyder (2004).
3
Examples include Robinson and Levy (1986), Berkowitz and Pritchard (1989), Delli Carpini and Keeter
(1989), Robinson and Davis (1990), Weaver and Drew (1993), and Arnold (2004). Mondak (1995) exploits
a quasi-experiment — a newspaper strike in Pittsburgh in 1992 that closed the city’s two major newspapers
for eight months. Unfortunately, he only has responses to vaguely worded questions about respondents’
knowledge of local politics.
5
Let MarketSharemc be newspaper m’s market share in county c. We define
Congruencecd =
M
X
MarketSharemc ReaderSharemd .
(2)
m=1
Congruencecd is the average ReaderSharemd , weighted by market share. It is a proxy for
how actively the average newspaper read in county c covers congressional district d, since
each newspaper’s incentive to cover district d is increasing in ReaderSharemd . Note that
since Congruence is defined using market shares, it is not dependent on the total newspaper
penetration in the county. The Appendix explains why Congruence is the best simple proxy
for news impact on voter information, given the relationship in equation 1.
To measure Congruence, we combine newspaper sales data with demographic data. Each
year, the Audit Bureau of Circulation collects data on each newspaper’s circulation in each
county, for almost all U.S. newspapers. We have this data for 1982 and for the period 19912004. We complemented this with county-circulation data for non-ABC newspapers for 2004
and 1991, and interpolated values in between those years. These were mainly smaller papers,
and the data was provided by SRDS.4 The U.S. Census Bureau collects data on the number
of people in each congressional district, by county. We have this data for the censuses of
1980, 1990 and 2000. For the years 1983-1990 when we do not have circulation data, we
interpolate Congruencecd . The formulas used for computing Congruence and ReaderShare
are shown in the Appendix.
To illustrate how Congruence varies, consider the case of Missouri, shown in Figure 2.
Newspapers are typically located in the larger cities and towns and sell to a markets around
these. The two largest cities in Missouri, Kansas City and St. Louis City both lie on the state
border and their markets cross this border. The third and fourth largest cities (disregarding
cities which are part of the Kansas City and St. Louis metropolitan areas) are Springfield
City, and Columbia City both located in the interior of the state and a congressional district.
Finally, the fifth largest city, St. Joseph City, is located on the state border but the area
across the border in Nebraska and Kansas are not densely populated.
The congressional district borders of Missouri are drawn with solid lines. First, take a
look at the congressional district to the northwest. In the north, St.Joseph News Press is
the dominant paper. We expect the St. Joseph News to mainly cover this congressional
district, since 93% of its readers live in this district. Therefore, Congruencecd is high in that
area. In the southern part of the district, people read the The Kansas City Star with 18%
of its readers in this district and thus, Congruencecd is much smaller. The Columbia Daily
4
On average there are about 10,900 observations each year in the ABC data, and about 500 observations
in the non-ABC data. There are about 3,000 counties in the U.S., so the average number of observations
per county in each year is slightly less than 4.
6
Tribune sells in the small appendix region to the southeast and only has 5% of its readers
in this district. As a result, Congruencecd is low.
The district in the southwest has high congruence. It has two major papers, The Joplin
Globe (with 80% of its sales in this district) and The Springfield News Leader (87%).
Note how the eastern counties of this district have high Congruencecd , whereas the adjoining county to the east of Springfield has a very low value. Similar strong differences in
Congruence between neighboring countries are found in the central and central-northeast
parts. These differences are due to the match between congressional districts and the markets
of a string of moderately sized papers: The Sedalia Democrat, The Columbia Daily Tribune,
The Post-Tribune/Capital Times, The Mexico Ledger and The Hannibal Courier-Post.
We also construct a measure for congruence of television markets, T V Congruence. To
identify radio and TV markets, we use the Designated Media Markets (DMAs). A DMA
is defined by Nielsen Media Research as all counties whose largest television viewing share
is given to stations of that same market area. 210 non-overlapping DMAs cover the entire
continental United States, Hawaii and parts of Alaska. Since we do not have viewership data
by county, we assume that everyone in a media market only watches broadcasts from that
market. For a county c in media market m, T V Congruencecd is thus defined as the population share of media market m that lies in district d. Figure 3 shows how T V Congruence is
constructed in Missouri. In the graph to the left, the solid lines are the congressional district
boundaries and the fillings identify the media markets. The graph to the right shows the
resulting T V Congruence.
A key identifying assumption in the empirical investigation is that Congruence is not
directly related to variables such as voters’ intrinsic interest in politics. We explore this
extensively below. For now, note that this seems a priori reasonable. The “economic geography” factors that determine media markets are generally quite different from the “political
geography” factors that determine congressional district boundaries.5 This is quite apparent
from Figure 3, showing both media markets and congressional districts. The match between
these economic and political geography concerns creates the haphazard patterns in Figures
2 and 3.
5
Congressional districts boundaries are drawn under the constraints that all districts in each state must
have the same population, constraints imposed by courts on racial considerations, partisan considerations,
the desire to protect incumbents, and so on. To satisfy the population requirement, congressional district
lines are redrawn after the population count of each Census. In our sample period, the congressional lines
were redrawn in 1992 based on the 1990 Census and in 2002, based on the 2000 Census.
Newspapers, on the other hand, typically locate in large cities, with strong demand for advertising and
news about city affairs. Their county sales depends strongly on the distance between the county and the
newspaper seat, and characteristics of the county residents such as their incomes, their level of education
and their age structure.
7
3
Newspaper Coverage
In this section, we examine the main force behind the results in this paper: that the number of
articles that a newspaper writes about a House representative is increasing in the estimated
fraction of the newspaper’s readers that live in the associated congressional district, the
ReaderShare.
Our data covers the period 1991-2002 (103rd-107th Congresses). The sample consists
of 161 newspapers which, on average, cover 385 districts for each congress.6 We used the
NewsLibrary.com web site, which employs a common search engine to search the on-line
archives of newspapers, for 142 newspapers. We supplemented this using Lexis/Nexis for 8
newspapers. In addition, we searched 11 newspapers’ web sites directly. In all, there are
4,216 observations in our sample, where each observation is a newspaper-district (newspaperrepresentative) pair in a given year.
We measure relative newspaper coverage as follows: Let qmdt be the number of articles
appearing in newspaper m during Congress t that contain
P both the name of the representative
from district d and the word “Congress.” 7 Let qmt = d qmdt , where the sum is taken over
all districts which contain at least 1% of newspaper m’s readers. Then, CoverSharemdt =
qmdt /qmt . This is the relative share of newspaper m.
The main independent variable of interest is ReaderSharemdt . (Note that we now use
t to denote time periods — each period is a Congress.) We include several other control
variables that are likely to affect the amount of coverage, including indicator variables for
party leaders (the Speaker of the House, Majority Leader, Minority Leader, Majority Whip,
and Minority Whip); seniority; an indicator variable for freshmen; an indicator for majority
party status; an indicator for out-of-state districts; variables for members in scandals; and a
variable indicating whether the representative sought higher office (governor or senator) or
received a higher appointment towards the end of the term. We also include some district
level demographics that might affect coverage, including income and the fraction of people
living in urban areas, the share old, the share foreign, and the share blue collar. Summary
statistics are in Table 1.
Figure 1(b) shows the basic pattern: a very strong, positive, and approximately linear relationship between CoverShare and ReaderShare. The bivariate regression line in
also shown in the figure. Table 2 shows that the relationship is essentially the same after
controlling for other factors that affect coverage.
Most of the controls have the effects that one would expect. Party leaders tend to receive
6
We do not have data on every newspaper for every year in the sample.
For simplicity, we use two calandar years for each Congress rather than the exact dates the Congress is
in session. Thus, for the 103rd Congress we use articles appearing January 1, 1991 through December 31,
for the 104th Congress we use articles appearing January 1, 1995 through December 31, 1996, and so on.
7
8
more coverage than other representatives. Members who are involved in scandals or who seek
higher office also receive more coverage. Three other findings are noteworthy. First, even
controlling for ReaderShare, newspapers exhibit an in-state bias, covering representatives
from their home state more heavily than out-of-state representatives. Second, even controlling for ReaderShare, urban representatives receive less coverage than others. This could
reflect the fact that urban areas have many powerful local politicians competing for scarce
newspaper space — mayors, police chiefs, school superintendents, and even city council members. Local governments provide a broader array of services in urban areas than in suburban
and rural areas, and citizens’ preferences in urban areas are more heterogeneous, so decisions
are more conflictual. As a result, citizens need to pay more attention to local politics, and,
since they are busy, they have less time to monitor their U.S. representatives. Finally, it
does not appear that newspapers are mainly interested in providing necessary information to
help voters vote in the next election (although that might be one consideration, of course).
For example, freshmen do not receive more coverage than other representatives, even though
they are new and relatively unknown to voters. Similarly, members who are retiring receive
just as much coverage — or even slightly more — as those running for re-election.
In quantitative terms, the share of the articles devoted to a representative increases by
about 60 percent when the share of the readership increases from 0 to 1. In a different
specification, we estimate the elasticity of the number of newspaper articles covering the
representative with respect to the reader share to 0.50 (s.e. 0.03). Since the typical newspaper
prints approximately 100 articles about a representative each year, this translates into about
50 more articles per year, although, as can be seen from the graph, there is a great deal of
variation for any given district readership share.
The very strong relationship between ReaderShare and press coverage of representatives
will drive our results. People who live in areas where Congruence (the market share weighted
ReaderShare) is high will, on average, be exposed to a considerably larger number of articles
about their congressman than people in other areas.
In order to determine what effects this may generate, it is useful to know what newspapers
actually report when they cover members of congress. Arnold (2004) studied newspaper
coverage of the House representatives in detail. In a representative sample of 25 newspapers,
covering equally many House representatives in the 103rd Congress, he found 8,000 news
stories mentioning the candidates. The articles were primarily news stories in reasonably
prominent places. Arnold found that the most commonly covered policy-related activities of
the congressmen are roll-call voting, advance position taking, and acting as local agents, that
is, working actively to acquire constituency benefits. Each of these three categories accounted
for about 1,200 news stories. If this is representative for our sample, then newspaper coverage
should inform voters of the representatives’ ideological positions, how well they act as local
agents, and, consequently, help voters rate their representatives.
9
4
Voter information
To hold their political representatives accountable, voters require information. Information
reduces the chances that voters make “mistakes” when casting their ballots, and increases
the electoral cost of politicians enacting poor policies, “shirking” and being corrupt. This
was the belief of the Framers, and is a standard feature of monitoring models of political
agency, see e.g. Persson and Tabellini (2000). A variety of models also predict that policy
becomes more distorted in favor of narrow interest groups when voters are less informed,
and that in a pluralistic world, policy outcomes favor groups with more informed members,
see, e.g., Grossman and Helpman (2001).
Information may also affect citizens’ welfare because it is useful in private decisions that
depend on government policies, or simply because voters dislike uncertainty. Bartels (1986)
and Alvarez (1997) present evidence that voters dislike candidates with uncertain policy
positions.8 Finally, easy access to information may increase political participation.
We will analyze five variables related to voter information. These variables are based on
survey data from the American National Election Studies (ANES) from 1982 to 2004. The
ANES is biannual, coinciding with congressional elections, and contains an average of about
1,800 respondents per year. The first two variables are based on direct knowledge questions.
The ANES asked respondents if they could name the candidates in the U.S. House races in
their district. We code the dummy variable NameRecall as one if a respondent could do
this, and zero otherwise. We code the variable NameRecognition as one if the respondent
could identify the incumbent from a list of the major party candidates for the U.S. House
in their district, and zero if they provided an incorrect answer. In our sample, 31% of the
respondents could correctly name at least one of the candidates, while a considerably larger
share, 92%, could recognize the incumbent’s name on a list; see Table 3.
The three final variables describe respondents’ willingness to describe or rate their House
representative. These types of variables have been used to proxy voter information, since
voters who know little or nothing about a representative are presumably less willing to describe or rate this representative.9 These variables are also directly related to accountability.
In order to hold representatives accountable, it helps to have an opinion on where they stand
ideologically and whether there is something to like or dislike about them. The ANES asked
respondents to place their House representative on a seven point ideological scale. We code
the variable IdeologicalRatingP rovided as one if the respondent provided a rating, and zero
if the respondent answered “don’t know” or “don’t recognize” the incumbent. The survey
also asked respondents to rate their feelings towards the incumbent on a scale from 0 to 100.
8
9
Although, see Berinsky and Lewis (2004) for a critique of some of this literature.
See, e.g., Delli Carpini and Keeter (1989).
10
We code the variable F eelingT hermometerP rovided as one if such a rating was provided,
and zero if the respondent answered “don’t know where to rate,” “can’t judge” or “doesn’t
recognize the incumbent’s name.” Finally, we code the variable LikesOrDislikesP rovided
as one if the respondent mentioned at least one thing he or she liked or disliked about the
incumbent House representative in their district, and zero if the respondent said that he/she
did not know anything about this candidate. In our sample, 70% were willing to describe
the incumbent’s ideological position, 82% provided a feeling thermometer rating, and 53%
mentioned something that they liked or disliked about the incumbent; see summary statistics
in Table 3. Our five variables are positively, but far from perfectly, correlated. For example,
80% of the respondents who can name a candidate mention something that they like or
dislike about the incumbent, as compared to 40% of those who cannot name any candidate.
The simple correlations are all larger than 0.2 and less than 0.4. Not all survey questions
were asked in all years, which accounts for the differences in the number of observations in
Table 3.
Our key independent variable is Congruence, and we use the following specification:
yit = γ Congruencecdt + xicdt δ + αt + αr + εi ,
where yi are our voter information variables, xicdt is a vector of control variables and αt a yearspecific fixed effect. Finally, αr is a fixed effect for each representative and three consecutive
terms, for example, a fixed effect for Linda Smith’s first three terms as representative for
Washington’s third congressional district, another for her next three terms, etc. Since the
dependent variable is dichotomous, the specification assumes a linear probability model. We
employ this because the linear probability model is consistent under weak assumptions, it
works well with fixed effects, and its coefficient estimates are simple to interpret. All reported
standard errors are heteroscedastic consistent.
Table 4 contains regression results. The columns contain different specifications, and the
rows contain results for our five dependent variables. The first column of Table 4 shows
the result of a regression of our information variables on Congruence, controlling only for
the representative- and year-specific fixed effects. Congruence is significantly and positively
correlated with all of these except NameRecognition. For this variable, the estimated coefficient on Congruence is of the same size as the standad errors in these regressions. Note
that the additional effect of Congrunce increasing NameRecognition is likely to be small,
since recognizing the representative’s name is fairly simple and over 90% do this. The other
estimated coefficients of Congruence are large. For example, a change from the lowest to
the highest values of Congruence is associated with a 28 percent increase in the probability
of correctly recalling a candidate’s name. This is about as large as the effect of changing a
respondent’s education from grade school to some college. We now try to establish that the
11
coefficient estimates in this column are good measures of the causal effect of Congruence
on voter knowledge. We extensively discuss identification concerns here. As many concerns
will arise again in the coming sections, we will frequently refer back to this discussion, for
example, for definitions of sets of control variables and specifications.
Adding controls Political knowledge is likely to be correlated with respondent characteristics such as education, income, and age. To distinguish the effect of the news market,
we include a large set of individual-specific controls in the vector xicdt . We include dummy
variables for whether the respondent’s party identification matches that of the incumbent,
for the respondent’s education (4 categories), income (5 categories), age (7 categories), gender, race, the number of years the respondent has lived in the community. These are 15
variables in total, after dropping one category of education, income and age to avoid perfect
collinearity.
We include seven county controls characterizing the county in which the respondent lives:
population (logged), average education levels (share with 1-11 years, share with 12 years,
and share with more than 12 years), per capita income (logged), share aged below 20, share
aged above 65, share black, and share female. We include eight representative controls:
the representative’s age, tenure, and dummy variables for whether the representative is a
freshman, belongs to the majority party, died or retired, serves on a powerful committee, or is
party leader, chair, or ranking member. There are three electoral-race controls: the closeness
of the race, measured by the negative absolute difference between the democratic share of the
vote and 0.5; a dummy variable for whether there was an incumbent running; and a dummy
variable for whether the incumbent ran unopposed. We include the following urbanism
controls: the degree of urbanism in the community where the respondent lives (3 dummy
variables), the share urban population, logged population density. There are four variables
in total, after dropping one category to avoid perfect collinearity. Finally, we include stateby-year fixed-effects, αsy , because growing states receive additional congressional districts
after redistricting, which typically changes Congruence, and state growth may be correlated
with our outcome variables.
The second column of Table 4 shows the result of regressions of our information variables on Congruence, adding all above mentioned controls. The estimated coefficients and
standard errors do not change considerably. For example, Congruence remains significantly
and positively correlated with knowledge about the candidates’ names. Most demographic
variables correlate with voter information in the expected direction. Older, better educated,
and higher-income respondents are more likely to be able to name a candidate, as are male,
white respondents. Moreover, respondents with the same party identification as the incumbent are significantly better at naming a candidate. More respondents can correctly name
12
a candidate when the incumbent representative is running opposed, and when the incumbent is young. Finally, voters are better able to name their representative when the county
population is larger. This is the only county-specific variable which is significant.
Within-race and redistricting specifications Our main identification concerns fall into
two categories: those related to representative characteristics and those related to population
or location characteristics. Regarding the first, more interesting and well known congressmen
attract more media coverage. If this additional coverage impacts newspaper sales, then it
will induce a spurious correlation between Congruence and voter knowledge. However,
the impact of coverage of the House representative on newspaper sales is likely to be very
modest. A typical newspaper prints about 100 stories per congressman and year; this can
be compared to the 1000 stories that are printed about the governor. Another apparent
representative-related concern is that certain types of congressmen self-select into highly
congruent districts. However, candidate selection is part of the media effects we wish to
study. A third representative-related concern is that congressmen who become increasingly
powerful manage to increase Congruence in their district during redistricting. We find no
correlation between being on powerful committees and increases in Congruence. Still, this
could be a potential concern.
We address these concerns by a within-congressional-race specification. By including a
district-by-year fixed-effect, we control for all factors that are constant within congressional
races, such as the characteristics of the incumbent and challenger, and the type of race. This
specification asks the question, are voters who live in counties with higher Congruence than
average in a certain district in a particular year better able to name their incumbent than
the average in that district in that year? We can ask this question because many districts
contain or overlap more than one county, so some voters (and survey respondents) in a given
district live in parts of the district with a high degree of Congruence, and others live in
parts with a low level of Congruence.
The results from the within-race specifications are show in Table 4, Columns III to IV.
The estimated coefficients on Congruence do not change to any considerable extent, nor do
the estimated standard errors.
The other category of concerns relates to population or location characteristics. The first
of these is that Congruence is lower in big cities. For example, New York is carved up into
a large number of congressional districts. Newspapers in the New York media market sell in
many districts and will not have a large share of their readers in any district. Other concerns
in this category are that Congruence may, for some reason, be correlated with unobserved
features that are also related to voter knowledge, such as unobserved income, education, and
age structure.
13
We address these concerns by a redistricting specification. This uses county fixed effects
and identifies Congruence effects through changes due to redistricting. We apply the 19822004 average sales of each newspaper in each county to all years. Consequently, Congruence
in a county-district only changes when the congressional district boundaries change due
to redistricting. For example, a county becomes more congruent if moved to a district
which their newspapers cover more (have larger sales in). In 1992, 1376 counties became
more congruent, 1532 became less congruent, and in 220 there was no change. The average
absolute change in Congruence was 0.1.
The results from the redistricting specifications are show in Table 4, Columns V and
VI. The estimated effects of Congruence retain their significance levels and do not change
considerably in magnitude. Finally, removing big cities from the sample and estimating the
specification in Column VI yields very similar results. For example, the estimated coefficient
on Congruence on NameRecall is 0.31 with a standard error of 0.07 (not shown in table).
Auxiliary regressions Another piece of evidence that Congruence is unlikely to be correlated with unobserved correlates of knowledge is that it is not significantly correlated with
those we do observe. As mentioned earlier, the estimated coefficients on Congruence in Table 4 are not very much affected by including the large set of control variables, even though
these variables are very significant. The reason is that our key variable, Congruence, is only
weakly correlated with the variables that affect individual respondents’ information levels.
We now investigate this directly by running auxiliary regressions of Congruence on all our
control variables. It would be worrying if, for example, demographic variables were correlated with Congruence. Note, however, that some control variables may be correlated with
Congruence. For example, in section 7 we will show that Congruence increases the support
for the incumbent. This means that more respondents may prefer the incumbent’s party in
highly congruent areas, and that incumbents in these areas may have longer tenure.
Table 5 shows the number of variables that are individually significant at the five-percent
level, with the total number of variables in that category in parenthesis, and the p-value of
a F-test for excluding all variables in that category.
Of the individual controls, the variable indicating that the respondent’s party matches
that of the incumbent is significant both in the baseline and the within-race specifications.
This correlation may be explained by a positive incumbency advantage effect of Congruence;
see below. As expected, Congruence is correlated with county characteristics, but not in
the Redistricting regression which includes county dummy variables. Similarly, Congruence
is correlated with some race and representative characteristics, but not in the within-race
specification where these partial correlations are not identified. In the redistricting specification, only 2 variables out of 40 are individually significant at the five percent level: the
14
dummy variables on the incumbent being unopposed (consistent with a positive incumbency
advantage effect) and living in a big city. To further guard against non-linear effects of urbanism from now on, we add 5 dummy variables for share urban and log population density
to the set of urbanism controls.
General knowledge effects As a final robustness check, we see whether Congruence
is related to general measures of voter knowledge. The congruence between a newspaper’s
market and a congressional district influences how much the newspaper writes about a particular congressman, but it should not influence how much the newspaper writes about many
other things, such as U.S. senators or the party that controls the U.S. House or the U.S.
Senate. Therefore, we replaced the dependent variable with knowledge questions not related
to a particular representative or district. Table 6 shows results from regressions using our
baseline specification. Columns I and II shows the results for the questions corresponding to
NameRecall and F eelingT hermometer but for senators. Column I shows a regression on
whether the respondent can name at least one Senator correctly (46 percent of the respondents were able to do this). In the second column, the dependent variable is a dummy for
whether the respondent was willing to place the incumbent senator on a feeling thermometer (91% where). We also check two other knowledge questions: whether the respondent
knows which party had a majority in the House before the election (56% correct), and which
party had a majority in the Senate (50% correct). The results show that Congruence is not
correlated with any of these measures of “non-district-specific” political knowledge. This increases our confidence that the positive correlations shown in Table 4, between Congruence
and district-specific information, are not spurious.
TV and radio Next, we investigate whether the effect mainly goes through newspapers or
other media. Newspaper markets and radio and TV markets are both centered around large
cities and towns. The two measures of Congruence and T V Congruence are fairly highly
correlated, with a correlation coefficient of .64.
We first evaluate how these two congruence measures explain differential media exposure.
In the surveys conducted between 1982 and 1994, the ANES asks respondents whether they
have read about the incumbent House representative in a newspaper or magazine (50% say
they have), or heard about him or her on TV (50%) or radio (25%). We constructed dummy
variables for positive responses to each of these three questions. Table 7 shows regressions of
these dummy variables on both Congruence and T V Congruence. The first column shows
that Congruence, but not T V Congruence, is significantly positively correlated with respondents saying that they have read about the incumbent in a newspaper or magazine. The next
two columns show that the opposite is true for radio and TV exposure. The estimated effects
15
are large. Congruence increasing from zero to one is associated with a 45% increase in the
share of respondents who say that they have read about the incumbent. This is consistent
with our newspaper content analysis, finding that an increase in ReaderShare from zero to
one is associated with a 50% increase in articles about the representative. A caveat regarding
these dependent variables is in place. We know from other studies that people have a hard
time remembering exactly where they read or heard something. For example, people report
seeing advertisements in areas where these advertisements have not been shown.
Interestingly, we find that T V Congruence is not significantly related to NameRecall,
while Congruence is; see the last column of Table 7. This is consistent with the view that
voters get most of their information about their local congressmen from newspapers rather
than television. This is plausible, since content analyses of television news suggest that
local television stations devote much less news coverage to local congressmen than do local
newspapers (Hess, 1991; Vinson, 2003; see also, the discussion in Arnold, 2004).
5
Representatives’ behavior in Congress
The results from the previous section indicate that voters in highly congruent districts are
better informed about their representatives in congress. These voters are better able to
name their representatives, more willing to place them ideologically, more willing to mention
something that they like or dislike about the representatives, and more willing to rate them
on a feeling thermometer. This is reasonable since the newspaper coverage of the representatives varies greatly with Congruence and the most commonly covered policy-related
activities of the congressmen are roll-call voting, advance position taking, and acting as local
agents (Arnold, 2004). We now expore whether news coverage influences the positions that
representatives take, especially on roll calls, and the effort they devote to acting as local
agents.
To provide a more concrete picture of what news coverage looks like, consider the case of
Representative Jack Brooks of Texas ninth district. Brooks represented a highly congruent
district, and was frequently in the press. In the 1994 election, his unpopular stance on
a crime bill was extensively covered. For example, the Houston Chronicle reports “Now,
Stockman [the challenger] is trying to portray Brooks as being soft on the issue because the
congressman supported the recently enacted crime bill, which bans assault weapons.”10 and
speculates about other provisons in the bill: “Was this a payoff by the Clinton administration
to buy Jack Brooks’ cooperation as he ignores his pro-gun constituents?”11 . The paper also
10
The Houston Chronicle, October 9, 1994, “Campaign ’94; Brooks feels right at home in the House;
Challengers saying incumbent’s 42 years in Congress are enough.”
11
The Houston Chronicle, August 21, 1994, “Did pork buy gun bill?”
16
discusses Brooks performance as a local agent: “If Brooks is so powerful, where’s the pork?
... Brooks can tick off a long list of public works projects he takes credit for bringing to the
district...”12 Brooks was serving on the Judiciary Committee, with limited direct influence
over grants. He was, however, actively lobbying for federal money to the district in frequent
witness testimonies before congressional hearings, as we discuss below. He lost the election
of 1994, and the defeat was partly attributed to his position on a crime bill.
Because they receive more press coverage, representatives from highly congruent areas
might have a greater incentive to engage in constituency service and pork-barrel politics. If
successful, they can look forward to local newspaper stories reporting on their success — free
advertising, from their point of view — and if unsuccessful they might find that reported as
well. Members representing districts with low values of Congruence will not receive much
free reporting of their activities, and even though the franking privilege reduces the costs
of promoting their records, self-promotion is probably viewed more skeptically than reporting by independent journalists. Similarly, representatives from highly congruent districts
might have greater incentives not to cast their roll-call votes against the interest of their
constituency. Doing so is more likely to generate publicity and electoral loss.
In this section, we investigate whether congressmen respond to these incentives. We
study whether representatives from highly congruent areas appear more as witnesses before congressional hearings, are over-represented on “constituency-oriented” committees and
under-represented on broad “policy-oriented” committees, and more frequently vote against
their party line.
5.1
Specification and identification
The level of analysis is now at the congressional district, and Congruence is now measured at this level. Suppose that M newspapers, indexed by m, sell in district d, and let
MarketSharemd be the market share of newspaper m in district d. Then,
Congruenced =
M
X
MarketSharemd ReaderSharemd .
m=1
At the district level, Congruenced has a mean of 0.45 and a standard deviation of 0.24, see
Table 8. Our unit of analysis will be congressional district by congress. We match witness
appearances, roll-call voting, and committee assignments each congress with Congruenced
in the coming election. For example, we match witness appearances in 2001 and 2002 with
Congruenced computed using the district lines of the 2002 election.
12
The Houston Chronicle, October 9, 1994, “Campaign ’94; Brooks feels right at home in the House;
Challengers saying incumbent’s 42 years in Congress are enough.”
17
We include various district controls: the percent urban and five dummy variables for
urbanness, log population density, and five dummy variables for log population density intervals, the log median income, the percent of, respectively, people aged 65 or older, military
population, people employed in farming, foreign born, and blue collar workers. We also
include the representative, and congressional race controls discussed in Section 4.
The identification concerns are similar to those discussed earlier, which is why we carefully
control for urban-rural differences. Since the analysis is now at the district level, we cannot
apply the within-race specification. We study the effects of redistricting, using district-fixed
effects. However, note that since the districts change with redistricting, the fixed effects will
not absorb all time-invariant heterogeneity. In sum, the identification is not as strong in this
section as elsewhere.
5.2
Witness appearances
One way for House representatives to advance their constituencies’ interests is by appearing
as witnesses at congressional hearings. This is a costly procedure and, on average, representatives only do it a few times per congress. To build the case that a certain project should
receive (or keep its) funding, the representative may have to gather data on and hire experts
discussing impacts on their district and the nation. Jack Brooks was a frequent witness,
appearing 14 times in 1994, of which 8 were before the appropriations committee. Consider,
for example, his testimony before the House Appropriations Committee, March 23, 1994.
“Mr. Chairman, thank you for this opportunity to appear today to urge
the subcommittee’s favorable consideration of the public works projects for the
ninth congressional district of Texas included in the fiscal year 1995 budget.
Mr. Chairman, I support funding for all of the projects included in the
budget proposal... All of these federal public works projects are critical to
the economic development of my district, which includes six major ports, and
many oil and chemical plant operations which compose one of the largest
industrial complexes in the United States. Ports along the Gulf coast are an
integral part of the economy of the nation... I have with me today a number of
persons vitally concerned with these projects. They are here to support these
statements and to answer any questions the subcommittee may have. I would
like to introduce them at this time...”
We have collected data on witness appearances by House representatives from 1982 to
2004. The variable W itness contains the number of appearances per congressional session by
each representative. Representatives stand witness before a congressional hearing on average
3.8 times per congress. However, there is considerable variation, as the standard deviation is
18
3.7 (see Table 8). The Appropriations and Ways and Means committees have a considerable
direct influence over spending and taxes, and witness appearances before these committees
may be more directly linked to influencing the allocation of funds. We therefore define the
variable W itnessBudget, containing the number of appearances before the Appropriations
or the Ways and Means Committees. Representatives stand witness before hearings in these
committees 1.7 times per congress, on average.
We want to investigate whether representatives from highly congruent districts stand
witness more often. First, we run a Poisson regression, controlling for our district controls
and state and year fixed effects. The results are shown in the first column of Table 9. The
coefficient on Congruence in the district is significant and positive. The point estimate
implies that an increase in Congruence from zero to one is associated with an average
increase of 46 percent in representative appearances. This implies 1.5 additional appearances
per congress. The next column shows the results from estimating the same specification using
a negative binomial regression (testing for goodness-of-fit rejects the Poisson specification).
The following Column (III) adds the representative and race controls. To that specification,
Column (IV) adds district fixed effects, and Column (V) replaces the district- and year- by
state-by-year-fixed effects.
The final two columns show the results from regressions on witness appearances before
the Appropriations and Ways and Means Committees only. The proportional effects are of
about the same size as witnessing before all committees. However, since the representatives
only witness before these committees an average 1.5 times per congress, the absolute effect
is smaller.
Representatives who are more closely monitored by the newspapers more often appear
as witnesses before committee hearings. The evidence presented in this section shows that
this correlation is significantly positive and robust across a wide range of specifications.
Surprisingly, even within the same congressional district, becoming more congruent over time
is associated with more witness appearances. The effects are also significant in appearances
before committees with a more direct say over the budget.
5.3
Committee choice
We now investigate whether representatives from congruent districts more frequently serve on
committees which are effective in promoting constituency interests. Congressional scholars
often divide the set of committees into four categories: constituency-oriented, policy-oriented,
prestige, and undesirable. The presumption is that representatives serving on constituencyoriented committees are better able to engage in constituency service and pork-barrel politics.
The opposite is true for representatives serving on policy-oriented committees, who instead
19
are better positioned to influence national policies.13
The committee assignments are influenced both by the individual representatives and
the party leadership. In the beginning weeks of the first session of every congress, new
congressmen submit requests, and returning congressmen may submit requests for second
assignments or transfers. Based on these requests, each party’s committee-on-committees
determines the committee seat allocation. The single factor regarded as most important in
distributing assignments to all but the prestige committees is whether a particular place will
help insure the reelection of the member in question (Masters, 1961). Other goals are policy
influence; see Fenno (1973) for a longer list of personal goals.
We now discuss our data and present the evidence. We have the committee assignments
of each representative from 1982 to 2004. While these are around 5,000 observations, the
number of independent observations is considerably smaller since many representatives remain on the same committee between congresses. In our sample period, there are about
1,000 representatives.
Operationally, we use Deering and Smith’s (1997) classification of constituency committees.14 We define the dummy variable DistribComm to equal one if more than 50 percent
of the representative’s committee assignments are on constituency-oriented committees, and
zero otherwise. This is true for about 28 percent of the representative-year observations.
We also constructing the variable P olicyComm which equals one if the representative serves
on the House Judiciary or the International Relations/Foreign Affairs Committees. This
variable is one for about 15 percent of the representative-year observations.
The committee assignments have been found to correlate with district demographics.
For example, Rhode and Shepsle (1973) find they are correlated with the population share
residing in urban areas, and the share of the labor force employed in the armed forces or in
agriculture. These variables are included in our district controls.
Our results from regressions of committee assignments on Congruence are presented in
Table 10. The first column shows that Congruenced is positively and significantly correlated with DistribComm, including only fixed effects for each year-state combination. In all
specifications, the standard errors are clustered by representative. However, when we add
13
Appropriations, Budget, Rules, and Ways and Means are designated as “prestige” committiees. See
Fenno (1973) for a detailed discussion and early classification. See, e.g., Baumgartner, Jones and MacLeod
(1997), Thorson, Glieden, and Lina (1999), Frisch and Kelly (2002) for recent papers using similar classifications.
14
Before the 104th Congress: Agriculture; Armed Services; Interior and Insular Affairs (Natural Resources
in the 103rd Congress); Merchant Marine and Fisheries; Public Works and Transportation; Science, Space,
and Technology; Small Business; and Veterans Affairs. In and after the 104th Congress, we designate the
following as constituency committees: Agriculture; National Security; Resources; Science; Small Business;
Transportation and Infrastructure; and Veterans Affairs.
20
the district, representative, and race controls, Congruenced loses significance; see Column
II. Columns III and IV use district fixed effects, with and without controls. Congruence
is insignificant in both these specifications. Columns V-VIII replace DistribComm with
P olicyComm as the dependent variable. Policy committee assignments are negatively correlated with Congruence, and significantly so in all specifications but one.
To sum up, we find weak evidence that Congruence affects committee assignments.
Congruence is mostly significantly negatively correlated with being on a policy committee,
but not consistently so across specifications. The effect of being on a distributive committee
is positive, but insignificant except in the most basic specification. This might reflect a low
statistical power because of the high persistence in assignments. The variable DistribComm
changes between congresses for only 7 percent of the representatives. A larger fraction, 22
percent, of representatives change their P olicyComm value between congresses.
5.4
Roll-call voting
Jack Brooks’ vote in favor of gun-control was in step with his party’s leadership, but not
with his constituency. In general, in determining whether to vote for or against a particular
bill, elected officials must balance the wishes of the overall electorate, specific constituencies
within the electorate, pressure from party leaders within the Congress, and their own ideology. We hypothesize that representatives in more congruent districts vote more in line with
constituency interests, because their constituencies are better able to monitor them.
We restrict the analysis to the ideological dimension of roll-call voting. There is some
empirical evidence that representatives’ voting is typically more ideologically extreme than
what their constituencies would prefer. A number of studies find a significant negative correlation between vote support and representatives’ “ideological extremism” in roll-call voting
(Democrats voting more to the left and Republicans voting more to the right); see e.g. CanesWrone et al. (2002). In a related vein, Ansolabehere et al. (2001) find that the ideological
gap between Republican and Democratic candidates is smaller in races that are expected to
be close. This presumably occurs because the competition for votes makes representatives
more prone to consider constituency interest over party and personal interests.
We test whether the ideological gap between Republican and Democratic representatives’
voting is smaller in highly congruent districts. To measure the ideological position of the
representatives, we use NOMINATE scores, constructed by Poole and Rosenthal (1997).
NOMINATE is a nonlinear latent variable model, somewhat akin to a factor analytic model,
used to estimate the underlying ideology that drives observed roll call behavior. Roughly
speaking, it assigns an ideological position to each legislator in order to correctly predict
as many roll-call votes as possible. The range of scores is [−1, 1], and scores are increasing
in conservativeness. The first dimension NOMINATE scores correlate well with other mea21
sures of ideology, such as ADA scores. We define the variable NominateScores as the first
dimension ideal point estimates from DW-NOMINATE.15
We have data on NOMINATE scores for House representatives for 1982-2002. Figure 4
plots these scores against the share of voters that supported the Republican candidate for
the House election in the congressional district. Scores from districts where Congruence is
above (below) the median are in grey (black). The solid lines show the predicted values from
separate linear regressions of Nominate scores on Democratic vote shares and a Democratwins dummy variable within the two sub-groups. The gap in voting is smaller in the highly
congruent districts. In other words, representatives are more moderate in their voting behavior in these districts. The same relationship is apparent in the “party loyalty” scores,
which is not surprising since ideologically extreme representatives tend to be those who vote
consistently with the party line. Recall from Figure 1(f) above that party loyalty (measured
as the percentage of times a member votes with the majority of the party leadership) is
inversely related to Congruence.
We now analyze these differences more carefully by regressing the NominateScores on
a dummy variable for the Democratic candidate winning the election, Congruence, and
the interaction of the two. The interaction allows the jump in NominateScores at 50%
Democratic votes to vary with Congruence. A positive interaction coefficient implies that
the negative jump at 50% is lower for more congruent districts. We also include fifth degree
polynomials in Democratic vote shares in the House and presidential elections to account
for any continuous influence on Nominate scores; using a specification similar to that in Lee,
Moretti and Butler (2004). The presidential vote share is interpolated for years in which
there was no presidential election. We limit the sample to elections where both parties had
at least a one percent vote share.
The results are presented in Table 11. The first column only contains only the above
mentioned variables. There is an estimated drop of -0.69 in NominateScores when a Democrat is elected instead of a Republican (at zero Congruence). The drop is -0.50 (-0.69+0.19)
when Congruence is one. The next column adds the district, race and representative characteristics discussed above. This causes the estimated coefficient on the interaction term to
drop to 0.17.
Because the drop may depend on district urbanism, we interact the dummy variable
for the Democratic candidate winning, not only with Congruence, but with the percent
of urban and five dummy variables for urbaneness, the log population density, and five
dummy variables for log population density intervals. The estimated coefficient on the
15
See Poole and Rosenthal (1997) for an exact definition of the statistical model underlying NOMINATE
and DW-NOMINATE, as well as discussion of the advantages of NOMINATE over other techniques and
indices.
22
interaction term is not much affected, see Column III. Because we interact with our urbanism
controls, the direct effect of the Democratic candidate winning does not have the same
interpretation as in Columns I and II. Column IV includes congressional-district and yearfixed effects. This specification identifies the coefficient on Democratic victory only through
districts where the party in power has changed. The final column adds the controls and
Democratic win dummy variable interacted with district urbanness. The interaction term
remains significantly positive.
We find that the ideological distance between Democratic and Republican representatives’ roll-call voting is about 25 percent lower (.17/.68=.25) in perfectly congruent than in
perfectly non-congruent areas. This is consistent with the hypothesis that representatives
moderate their behavior in highly congruent districts because the better monitoring of the
newspapers increases the electoral costs of ideologically extreme or highly partisan roll-call
voting.
6
Policy
We have argued that representatives that are closely monitored by local newspapers have
a greater incentive to provide services to their constituency. We have also provided some
evidence that representatives react to these incentives, they appear more as witnesses before
congressional hearings, perhaps sit less on policy committees, and display less extreme rollcall voting patterns. If the representatives’ efforts in Congress are successful, then we should
observe more federal funds flowing into highly congruent districts, and to more congruent
counties within districts.
To test this, we assembled data on federal expenditure allocations across counties from
the Consolidated Federal Funds Report. The expenditures we analyze include grants, procurement contracts, salaries and wages, and direct payments for retirement and disability,
and other direct payments. We exclude loans, insurance and social security payments.16
The expenditures are in constant 2000 dollar values. The total value of the expenditures
we study is about $ 2,700 per capita, or about 10 percent of U.S. GDP. We have this data
yearly for the period 1983-2004. However, we collapse the data by Congressional session,
and study mean annual expenditures per Congress. We define the variable Spending as the
(log of) this mean annual spending per Congress.
Since the unit of analysis is now counties and some counties cross district boundaries,
we average the variables containing district controls. We compute the population weighted
16
The loans data does not report the value of the loan subsidy but only the value of the loan, the insurance
data reports obligations and not payments, and the social security data is excluded because representatives
are unlikely to affect its distribution.
23
average Congruencecd in a county
Congruencec =
X ncd
nc
Congruencecd ,
where ncd /nc is the share of the population of county c that lives in district d. Similarly, we
compute the population weighted averages of district level variables such as representative
seniority, committee assignments, etc., and use these in the regressions.
Table 12 shows the results of a set of regressions of Spending on Congruencec . We
apply the baseline, within-race and redistricting specification with and without of our sets of
county, urbanism, representative, and race controls. We also include a dummy variable for
the county containing the state capital. Once more, it is important to control for the urban
dimension since it correlates with both federal spending and Congruence. In the withinrace specification, we only look at counties which lie entirely in one district. We also add a
“within neighbor-counties” specification where each county that lies along a district border
has been matched with the closest county on the other side of the border. This specification
includes dummy variables for each such neighbor pair.
The coefficient on Congruence is significantly positive in all specifications that use crosssectional variation (which is all but the last two). Comparing counties within the same
district and year, significantly more funds were spent in more Congruent counties. Comparing neighboring counties on different sides of district borders, significantly more funds were
spent in the more congruent neighbor. The estimated effect is consistently about 0.1 which
implies that a shift in Congruence from zero to one is associated with a 10-percent increase
in spending.
In the redistricting regression, the coefficient on Congruencec is only significant at the
10-percent level after adding controls. The size of the coefficient also drops to 3-5 percent.
The reason, we believe, is that there is a very high inertia in spending. Much of the funding
that the government spends in a given year was authorized in previous sessions of Congress
as the total money allocated is the sum of continuing awards still in existence as well as
new funds being awarded that year. For this reason, the stock of spending is less related to
congressional action in any given year. It may be possible to address this issue by using data
that identifies new awards by year, such the U.S. Domestic Assistance Programs Database
used by Bickers and Stein. We leave this for future work.
In sum, we find strong evidence that media coverage substantially influences federal
spending. This is the last link in the accountability chain. Our results on news coverage,
voter information, politicians’ actions and policy consistently support the idea that an active
media coverage of politics increases electoral accountability.
24
7
Extensions:
Voter turnout and incumbency advantage
We next turn to two important additional issues: the effects of press coverage on voter
turnout and the incumbency advantage.
7.1
Voter turnout
For most Americans, voting is the main form of political participation. This participation
is, in the view of many observers, beneficial per se. A high turnout in fair elections gives
legitimacy to public officials and their decisions. Voting is also an effective means of holding
elected officials accountable for their decisions and behavior in office. Voter abstention
may signal that citizens are dissatisfied with their political institutions, or lack a sense of
efficacy and confidence in government, and habitual abstention may further erode citizens’
satisfaction and confidence.
It is clear that Congruence could influence political participation. We have seen that
it affects voters’ information, and numerous surveys find a strong and positive correlation
between citizens’ information levels and political participation.17 While causation is more
difficult to establish, several studies are suggestive. The strong correlation between education
and turnout may also be interpreted as evidence that decreasing the cost of information
increases participation.18
The analysis of voter turnout in House elections is complicated by the fact that people
vote for multiple offices (and sometimes also controversial and important referenda). The
decision to go to the voting booth depends on all these elections. One way of dealing with
this is to control for turnout for the most important office. This specification analyzes “rolloff.” That is, given that a voter has come to vote for, say, the office of president, how likely
is it that he or she will not vote for Congress?19
To analyze voter turnout, we use data on votes in congressional elections, by county and
congressional district , for 1984-1992 and 1998-2004.20 Voter turnout in the presidential
17
See, e.g., Verba and Nie (1972).
The underlying assumption is that education reduces the costs of acquiring information. See, e.g.,
Wolfinger and Rosenstone (1980) for evidence of the correlation.
19
Another possibility is to study elections when there was no presidential, gubernatorial or senatorial
contest on the ballot, so the House race was arguably the “top of the ticket.” Unfortunately, we cannot do
this because there are two few such cases, from a small number of states.
20
We could also use the self-reported turnout data in the ANES. However, it is well known that respondents
grossly overstate their propensity to vote, and also overstate their propensity to vote for winning candidates
and incumbents (at least in U.S. House races). For this reason, we prefer the actual voting data. When we
18
25
election will be used as a control, and we have data on presidential election returns by
county and district for the 1984, 1988 and 2000 elections. We drop uncontested elections
and observations where the reported election turnout is larger than the total population.
In our sample, the average share of the population who voted in the House election was
36 percent. The turnout rate is substantially higher in presidential election years (about 40
percent) than in other years (about 30 percent). When we have data on the number of votes
cast both in congressional and presidential elections, the average turnout is 1.2 percent lower
in congressional in presidential elections.
We regress the turnout in House elections on Congruence, controlling for turnout in the
presidential election. The results are shown in Table 14. The coefficient on Congruence is
highly significant in the baseline and within-race specifications. In the redistricting specification, the standard errors are large and the estimated effect is not significant. Apparently,
there is not enough time-series variation to identify effects.
In the baseline specification, an increase in Congruence from 0 to 1 is associated with
a 0.8 percent increase in House election turnout (as share of the population). This is twothirds of the average 1.2 percent roll-off if presidential election years. While this is a large
percentage decline in the “roll off”, the absolute increase in turnout is not large. It is possible
that the effect is larger in elections when the House election is at the “top of the ticket.”
However, running the same regression as in Table 14 for this sub-sample yields an estimated
effect in the baseline regression of 0.8 with a standard error of 0.6. So it is unlikely that the
impact is larger than, say, 2 percent.
Given earlier large estimated effects of media access on turnout, these moderate effects
are somewhat surprising. Strömberg (2004) estimates that the increasing use of radio in
1920-1940 led to an increase in votes per capita of about 5.5 percent in gubernatorial elections. More closely related to our study, Gentzkow (2006) estimates that the increasing
use of television in 1950-1970 reduced House turnout in the non-South by 5.6 percent in
off-year elections and 3.1 percent in presidential elections. The reason is, he argues, that
people stopped reading newspapers and listening to the radio and instead turned to the uninformative TV. Our estimates suggest that if local newspapers completely stopped covering
House politics, then turnout would drop by less than one percent in presidential-election
years. One possible explanation for the differences is that the introduction of radio and TV
affected voter turnout for all offices, while Congruence only affects House turnout.
analyze the ANES we find significant and even larger effects of Congruence both on voter turnout and the
incumbency advantage.
26
7.2
Incumbency advantage
The electoral advantage of incumbents is one of the most studied issues in U.S. congressional
politics. The dramatic increase in the incumbency advantage since the 1950s has sparked
a discussion of its origins, and electoral, democratic and policy consequences. A central
question in this research is why this advantage has been growing so rapidly since the 1950s.
Researchers have proposed a variety of potential explanations including: the electorate has
become more ideologically polarized, incumbents have been able to create safe districts by
redistricting, incumbents control increasingly valuable resources for constituency service, and
incumbents have received increasing media exposure. We investigate a question relating to
this last point. Specifically, we study whether people in areas with more media coverage of
congressional politics are more likely to vote for the incumbent.
It is not obvious that voters with more information will be more supportive of incumbents.
However, if pure name recall acts as a signal of candidate quality, then media coverage may
increase electoral support.21 The incumbency advantage may also increase if media coverage
of incumbents is biased towards positive news, or if media provides more precise information
about incumbents and voters are risk averse. Finally, changes in representatives’ behavior
induced by media coverage may influence voter support.
Using the ANES survey data, we first analyze whether newspapers differentially inform
voters about challengers and incumbents. This can be done since the ANES records separately whether respondents correctly recalled the names of both the incumbent and the
challenger running in their district. The ANES also records separately whether respondents
report having read about the incumbent or the challenger. We created dummy variables for
each of these four items. Incumbents are better known than challengers, and more people
read about them. In contested elections with an incumbent running for re-election, about
50% of the respondents say that they have read about the incumbent, compared to 15% for
the challenger; and about 30% can correctly name the incumbent, compared to 10% for the
challenger.
What are the differential impacts of Congruence? Table 15 shows the results from
regressions of the “read about” and name recall dummy variables on Congruence. Higher
Congruence significantly increases both incumbent and challenger name recall. However, the
effects are larger for incumbents than for challengers (0.33 as compared to 0.12). Moreover,
while higher Congruence also significantly increases the probability of reading about the
incumbent, it appears to have little effect on reading about challengers.
We now turn to voting data to see whether incumbents receive more vote support in highly
21
For example, good representatives may be more active in Congress and hence receive more media coverage. See Nelson (1970) or Kihlstrom and Riordan (1984) for a similar argument for consumer product
advertisements.
27
congruent areas. Following the previous literature, we measure the incumbency advantage
by regressing the Democratic vote share on an indicator variable for whether the incumbent
House representative is a Democrat, controlling for other factors that determine the election
outcome. To test whether incumbents are more advantaged in highly congruent areas, we
add the interaction between the dummy variable for Democratic incumbent and Congruence.
We demean the incumbency advantage and Congruence variables before interacting them,
so that the coefficients on the main effects show the impact of these variables measured at
their sample means. The specification is of the form:
0
vhcdt = β 1 Idt + β 2 Idt
Congruence0cdt + β 3 Congruencecdt + normalvotecdt
0 0
+xcdt δ + Idt
xcdt δ + αst + εcdt ,
where vhcdt is the Democratic share of the two party vote in county c in district d in the
House election at time t, and Idt is a dummy variable for whether the incumbent in district d
0
was a Democrat. The prime superscript denotes demeaned variables, for example, Idt
equals
Idt minus the sample mean of Idt . As controls, xicdt , we include the race, representative,
county, and urbanism controls discussed in section 4. We also interact the controls with
the incumbency dummy variable to allow the incumbency effect to vary with the degree of
urbanness, etc.
We use two sets of variables proxying normalvotecdt , the constituencies’ underlying partisan preferences (e.g., Levitt and Wolfram, 1997, and Gelman and King 1990). The Levitt
and Wolfram (1997) specification uses a district-by-county fixed effect for each districting period. The other specifications include the lagged Democratic vote in district d and a dummy
variable for the Democrats being the incumbent party. Both specifications include the Democratic share of the two-party vote in the presidential election in the county-district part.
Since our unit of observation is county by district by election, we can both use redistricting
and within-district specifications. The results are show in Table 16.
We estimate an average personal incumbency effect of 8-13 percent, which is similar to
previous estimates in the literature. In areas where we expect the press to cover congressional politics more, the incumbency advantage is about 1 percentage point larger. This is
true whether or not we include controls, and whether we use the baseline, within-race, or
redistricting specifications.
The incumbency advantage in House elections has risen from less than two percent in
the 1940s to about eight percent today. Our estimates indicate that a relatively small share
of this can be explained by changes in the media market. Given the large estimated effects
on voter information, it also seems unlikely that increased voter information is a major
determinant of the rise in the incumbency advantage.
28
8
Discussion and Conclusion
All is not safe, even where the press is free and every person is able to read. The press must
also have economic incentives to cover politics. Those incentives might be muted by a poor
fit between media markets and political districts. Stated from the positive side, we find that
a good fit (high Congruence) between newspaper markets and congressional districts leads
to more press coverage of the local congressmen. In areas where coverage is high for this
reason, voters are better informed about their congressmen and more likely to participate
in elections. Politicians respond to the increased media coverage by more actively pursuing
their constituencies’ interests. They moderate their partisan voting, more frequently stand
witness before committee hearings, and (perhaps) more frequently work on constituency
committees. As a result, we see more federal money flowing into more congruent districts.
Quantitatively, the effects of news coverage on voter information, politicians’ actions
and policy are large. We estimate that increasing Congruence from zero to one increases
the number of articles written about the representatives by 50 percent, it increases voters’
ability to correctly name the candidates in the House race by 30 percent and their willingness
to describe and rate the representative by 20-30 percent, it increases the representatives’
witness appearances by 46 percent, and reduces the gap between Republican and Democratic
ideological differences in voting by 25 percent. Finally, it increases federal spending per
capita by 10 percent. A one standard deviation increase in Congruence (0.45) induces
effects that are slightly less than half this size. So the variation in Congruence induces a
significant variation in all these outcomes. Effects on voter participation and the incumbency
advantage are significant, but not large.
Our findings support the idea that press coverage is important for electoral accountability.
Voters need information to keep politicians accountable and the press delivers this information. This was the belief of the Framers, and, that information improves accountability is
also a standard feature of political agency models; see Persson and Tabellini (2000) or Grossman and Helpman (2001). The U.S. House districts constitute a suitable testing ground for
this idea. While the districts are of similar size and political competence, press coverage
varies (we argue) exogenously with the Congruence between media markets and congressional districts. Moreover, the importance of the U.S. House makes the welfare consequences
of public action or inaction of high order.
Our results are helpful in understanding the political consequences of trends in media
markets. At the start of the twentieth century, every large and medium-sized U.S. city and
even many small towns had at least one newspaper, geared to local political needs. The
number of newspapers has shrunk so that many communities no longer have a paper of their
own. Newspapers have merged and been replaced by broadcast media, serving even larger
regions. This trend is likely to decrease the congruence at local levels. This might not only
29
affect the congruence of congressional districts, but also that of municipalities, counties,
and other local government units. Our results suggest that this is likely to reduce voter
information, political participation, and political accountability.
The internet may influence this trend. Already at this stage, more than 1,300 domestic
daily newspapers have a web presence, and 15 percent of all American adults say that the
internet was the primary source of campaign news during the 2006 election (compared to
34% for newspapers).22 The effect on information about local politics is unclear. On the one
hand, there is a low fixed cost of starting an internet outlet, basically no space constraint,
and good search technologies. This should increase the number of viewable relevant articles.
On the other hand, most people do probably not search intensively for information about
local politics, rather they get this information as a by-product. In addition, the cost of
gathering, editing and writing of news is the same for print and online editions, and these
costs make media write more about issues that concern larger shares of their readership.
Moreover, since the potential internet audience is global, local reader shares may decrease.23
Because of the large effects, Congruence should be an important factor to consider
in contemplating the regulation of the press to improve voter information and political
accountability. It may be even more important than, say, the degree of competition between
local newspapers. There exists an externality argument for why congruence may be too
low. Better informed voters induce politicians to work more for their constituencies, which
benefits all voters. However, individual voters do not fully internalize these benefits and,
consequently, their willingness to pay for political information is low (see Downs, 1957).
This causes media markets to conform too little to political districts.
Note that our results may generalize to media coverage of issues. In the same way that
a representative works more for geographical constituencies that are better informed, they
may also work more for better informed issue-defined constituencies (such as environmental
groups, or people who care strongly about taxes). If our results generalize in this way, they
have wide-ranging implications. The media coverage of issues varies greatly across issues in
a systematic way. This could create a systematic policy bias disfavoring minority groups and
groups not valuable for advertisers, see Strömberg (1999).
Finally, as an application, the strong “personal vote” in U.S. elections might in part be
due to the local nature of much newspaper coverage, and could therefore be effected by these
22
PEW Research Center (2007), “Election 2006 Online.”
The UK’s The Guardian newspaper receives 78 per cent of its web readers from overseas (Mayes, 2004),
compared to 30 per cent for The New York Times and 17 per cent for The Washington Post (New York Times,
2005; Tedeschi, 2004). The concentration in online media use is also very high. Although thousands of sites
have contain information on any given topic, only a few sites account for most of the traffic. For example,
three old media-related sites–MSNBC, CNN, and The New York Times–account for three-quarters of
traffic online for U.S. news (Barnhurst, 2001).
23
30
trends. Today, relatively few voters in the U.S. read a “national” newspaper; rather, almost
everyone who reads a newspaper reads a local paper. In 2003, the combined circulation of the
three largest newspapers — The New York Times, The Wall Street Journal, and USA Today
— was 6 million out of a total circulation in the U.S. of 50 million, or just 12 percent. In
many other countries, the opposite is true. Citizens in these countries are therefore unlikely
to encounter much news about their local politicians. This is even the case in countries with
geographically defined constituencies, such as the U.K.
31
9
References
Ahrend, Rudiger (2002) “Press Freedom, Human Capital, and Corruption” , Working Paper
no. 2002-11, DELTA.
Alvarez, R. Michael (1997) “Information and Elections” , Ann Arbor, MI: University of
Michigan Press.
Ansolabehere, Stephen, James M. Snyder, Jr., and Charles Stewart, III (2001), “Candidate Positioning in U.S. House Elections” , American Journal of Political Science 45
(January), 136—59.
Ansolabehere, Stephen, Eric C. Snowberg, James M. Snyder, Jr. (2006) “Television and the
Incumbency Advantage in U.S. Elections”, Legislative Studies Quarterly 31(4), 469-490.
Arnold, R. Douglas (2004), “Congress, the Press, and Political Accountability”, Princeton,
NJ: Princeton University Press.
Barnhurst, Kevin G., (2002), “News Geography & Monopoly: the form of reports on U.S.
newspaper internet sites”, Journalism Studies, 4(3), 477—489.
Bartels, Larry M., (1986), “Issue Voting Under Uncertainty: An Empirical Test”, American
Journal of Political Science 30, 709-728.
Baumgartner, Frank R., Bryan D. Jones, and Michael C. MacLeod, (1997), “New Issues and
Old Committees Jurisdictional Change in Congress, 1947—93.”Unpublished manuscript,
Pennsylvania State University.
Berinsky, Adam J., and Jeffrey B. Lewis, (2004), “Testing the Assumption of Quadratic Preferences: Voter Choice and Candidate Uncertainty in Spatial Models of Issue Voting.”
Unpublished manuscript, Massachusetts Institute of Technology.
Berkowitz, Dan, and David Pritchard, (1989), “Political Knowledge and Communication
Resources”, Journalism Quarterly, 66, 697-701.
Besley, Timothy, and Robin Burgess, (2002), “The Political Economy of Government Responsiveness: Theory and Evidence from India”, Quarterly Journal of Economics 117,
1415-1452.
Besley, Timothy, and Andrea Prat, (2006), “Handcuffs for the Grabbing Hand? Media
Capture and Government Accountability”, The American Economic Review 96(3), 720736.
Brunetti, Aymo, and Beatrice Weder (2003), “A Free Press Is Bad News for Corruption”,
Journal of Public Economics 87, 1801-1824.
Campbell, James E., John R. Alford and Keith Henry (1984) “Television Markets and Congressional Elections”, Legislative Studies Quarterly 9, 665-678.
Canes-Wrone, Brandice, David W. Brady and John F. Cogan, (2002) “Out of Step, Out
of Office: Electoral Accountability and House Members’ Voting”, American Political
Science Review 96(1), 127-140.
32
Cohen, Marty, Hans Noel, and John Zaller (n.d.) “Local News and the Quality Of Congressional Representation”, Unpublished manuscript, UCLA.
Delli Carpini, Michael X., and Scott Keeter (1989) “What Americans Know about Politics
and Why It Matters”, New Haven, CT, Yale University Press.
Deering, Christopher J., and Steven S. Smith (1997) “Committees in Congress”, Washington,
DC, Congressional Quarterly Press.
Djankov, Simeon, Caralee McLiesh, Tatiana Nenova, and Andrei Shleifer (2003), “Who Owns
the Media?”, Journal of Law and Economics 46, 341-383.
Downs, Anthony (1957) “An Economic Theory of Democracy”, New York, Harper Collins.
Fenno, Richard F., Jr. (1973) “Congressmen in Committees”, Boston, Little, Brown.
Frisch, Scott A., and Sean Q. Kelly (2002) “House Committee Assignments and Public
Policy: Committee Theories Reconsidered”, Unpublished manuscript, California State
University at Bakersfield.
Gelman, Andrew and Gary King, (1990) “Estimating Incumbency Advantage without Bias”,
American Journal of Political Science 34(4), 1142-1164.
Gentzkow, Matthew (2006), “Television and Voter Turnout”, The Quarterly Journal of Economics 121(3), 931-972.
Grossman, Gene, and Elhanan Helpman (2001) “Special Interest Politics”, Cambridge, MA,
MIT Press.
Hess, Stephen (1991) “Live From Capitol Hill!”, Washington, DC, Brookings Institution.
Kihlstrom, Richard E., and Michael H. Riordan (1984) “Advertising as a Signal”, Journal
of Political Economy, 92(3), 427-450.
Lee, David S., Enrico Moretti and Matthew Butler (2004) “Do Voters Affect or Elect Policies?
Evidence from the U.S. House”, Quarterly Journal of Economics 119(3), 807-859.
Levitt, Steven D. and Catherine D. Wolfram, (1997) “Decomposing the Sources of Incumbency Advantage in the U. S. House”, Legislative Studies Quarterly 22(1), 45-60.
Levy, Dena, and Peverill Squire (2000) “Television Markets and the Competitiveness of U.S.
House Elections”, Legislative Studies Quarterly 25, 313-325.
Masters, Nicholas A. (1961) “Committee Assignments in the House of Representatives”, The
American Political Science Review 55(2), 345-357.
Mayer, William G. (1993) “The Polls: Trends in Media Usage”, Public Opinion Quarterly
57, 593-611.
McCarty, Nolan M., Keith T. Poole, and Howard Rosenthal (1997), “Income Redistribution
and the Realignment of American Politics”, Washington, AEI Press.
Mondak, Jeffrey J. (1995) “Newspapers and Political Awareness”, American Journal of Political Science 39, 513-527.
Nelson, Phillip (1974) ”Advertising as Information”, Journal of Political Economy 82, 729754.
33
Niemi, Richard G., Lynda W. Powell, and Patricia L. Bicknell (1986), “The Effects of Congruity between Community and District on Salience of U.S. House Candidates”, Legislative Studies Quarterly 11, 187-201.
Poole, Keith T., and Howard Rosenthal (1997) “Congress: A political-economic history of
roll call voting.”, New York: Oxford University Press, 1997
Persson, Torsten and Guido Tabellini (2000) “Political Economics: Explaining Economic
Policy”, MIT Press.
Rohde, David W., and Kenneth A. Shepsle (1973) “Democratic Committee Assignments
in the House of Representatives: Strategic Aspects of a Social Choice Process”, The
American Political Science Review 67(3), 889-905.
Robinson, John P., and Mark Dennis K. Davis (1990) “Television News and the Informed
Public: An Information-Processing Approach.”Journal of Communication 40, 106-119.
Robinson, John P., and Mark K. Levy (1986) “The Main Source: Learning from Television
News”, Beverly Hills, CA, Sage.
Stewart, Charles, III and Mark Reynolds. 1990. “Television Markets and U.S. Senate
Elections.”Legislative Studies Quarterly 15:495-523.
Strömberg, David (1999) “The Political Economy of Public Spending”, Ph. D. Dissertation,
Princeton University.
Strömberg, David (2004) “Radio’s Impact on Public Spending”, Quarterly Journal of Economics 119(1), 189-221.
Thorson, Greg, Lynnae Glieden, and Jason Lina (1999) “Swinging Left: An Examination of
the Change in Committee Compositions in the U.S. House during the Reform Period”,
Unpublished manuscript, University of Minnesota at Morris.
Verba, Sidney, and Norman H. Nie (1972) “Participation in America”, Chicago, University
of Chicago Press.
Vinson, C. Danielle (2003) “Local Media Coverage of Congress and Its Members: Through
Local Eyes”, Cresskill NJ, Hampton Press.
Weaver, David, and Dan Drew (1993) “Voter Learning in the 1990 Off-Year Election: Did
the Media Matter?”Journalism Quarterly 70, 356-368.
Welch, Susan (1978) “Election to the Legislature: Competition and Turnout”, In John C.
Comber and James B. Johnson, eds., Nonpartisanship in the Legislative Process: Essays
on the Nebraska Legislature, Washington, DC, University Press of America.
Wolfinger, Raymond E., and Steven J. Rosenstone (1980) “Who Votes?”, New Haven, CT,
Yale University Press.
34
Appendix
Computation of Congruence
Let qmd denote the number of articles that newspaper m writes about House Representative from district d. Assume that
qmd = αReaderSharemd .
Next, assume that the share who reads any article is proportional to the share of households
who buy the paper (that we call ReaderP enetration). This assumption is standard in the
advertising literature. Let pmc be the probability that a random person in county c reads
any particular article in newspaper m. We assume that
pmc = hReaderP enetrationmc .
Typically, the share of buyers who read an article, h, will depend on the article’s size, location
and attractiveness as well as features of the newspaper. Since we do not have data on these
features, we simply use the average probability h.
Now, if newspaper m writes qmd articles about a congressman, and each time the probability that a person reads the article is pmc , then the expected number of articles that
this person will read about the congressman in this newspaper is pmc qmd . Suppose that M
newspapers sell in county c and each newspaper m prints qmd stories about the congressman.
Then the total expected number of articles that a random person in county c will read about
his or her congressman will be24
M
X
m=1
pmc qmd = αh
M
X
ReaderP enetrationmc ReaderSharemd = αhExposurecd .
m=1
To address a potential endogeneity problem, we need to rewrite the equations before
estimation. The Exposure variables are mixtures of two concepts: (i) “congruence,” i.e., the
extent to which a district’s boundaries match the boundaries of newspaper markets; and (ii)
readership, i.e., overall newspaper penetration in a district or county. It is more than likely
that newspaper penetration is correlated with unobserved factors which are correlated voter
behavior, since people who buy newspapers are richer, better-educated, more interested in
politics, etc.
24
Not only the expected number, but also the distribution of articles read could matter for average learning.
The MEDIAC model of advertisement impact discusses how to account for this. The whole distribution
matters if learning is not linear in the number of articles read. Marginal effects may increase initially since
the first articles generate interest, so called "wearin". At higher levels, the effect on learning is decreasing
in the number of articles read (Pechmann and Stewart, 1989).
35
To avoid this problem, we define
Exposurecd
,
m ReaderP enetrationmc
or equivalently,
Congruencecd = P
Congruencecd =
M
X
MarketSharemc ReaderSharemd .
(3)
m=1
Congruencecd is the average ReaderSharemd , weighted by market-share. It is a proxy for
how actively the average newspaper read in county c covers congressional district d, since
each newspaper’s incentive to cover district d is increasing in ReaderSharemd . Intuitively,
Congruencecd is the value Exposurecd would take if everyone in county c reads exactly one
newspaper — this would be the average share of newspapers’ markets which lie in district
d. Congruence varies between zero and one. Note that since Congruence is defined using
market shares, it is not dependent on the total newspaper penetration in the county.
Empirically, we compute Congruence as follows. We use the newspaper sales data to
estimate the number of people in each district that read each newspaper. (Although we
estimate this separately for each year in our sample, we suppress time subscripts in what
follows.) P
Let ncd be the number of people who live in county c and congressional district
P d,
let nc = d ncd be the total number of people who live in county c, and let nd = c ncd
be the total number of people who live in district
P d. Let xmc be the number of copies of
newspaper m sold in county c, and let xm = c xmc be the total number of copies sold of
newspaper m. Assume that the number of copies per capita of newspaper m is the same in
the part of county c that live in district d as for the county as a whole (we do not have data
on newspaper sales for sub-county units). Then the number of copies of newspaper m sold
in the part of county c that lies in district d is approximately
xmc
xmcd =
ncd .
(4)
nc
Also, the number of copies of newspaper m sold in district d is approximately xmd =
P
c xmcd . These variables can then be used to compute MarketSharemd , ReaderSharemd
and Congruencecd using the definitions
xmc
,
MarketSharemc = P
m0 xm0c
and
xmd
.
xm
Combining these, we simply use equation (2) to compute Congruencecd .
ReaderSharemd =
36
Table 1. Newspaper data summary statistics
Variable
Obs
Mean Std. Dev .
4213
0.15
0.15
CoverSharemd
4213
0.15
0.20
ReaderSharemd
4213
0.01
0.14
Scandal
4213
0.02
0.13
Party leader
4213
0.03
0.17
Higher office
4213
0.26
0.16
Close race
4213
0.14
0.35
Freshman
4213
0.09
0.28
Retired
4213
0.13
0.34
Out of state
4213
0.68
0.30
Pct. urban
Min
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
Max
1.00
0.99
2.00
1.00
1.00
0.50
1.00
1.00
1.00
1.00
Table 2: Newspaper Coverage of U.S. House Members, 1991-2000
Dependent Variable: CoverSharemd
I
II
ReaderSharemd
0.64
0.62
(0.02)*** (0.02)***
0.10
Party leader
(0.02)***
0.04
Scandal
(0.01)***
0.08
Higher_office (ran or appointed)
(0.01)***
-0.03
Out_of_state
(0.01)***
-0.09
Pct. urban
(0.01)***
0.04
Close_race
(0.01)***
0.00
Freshman
(0.00)
0.02
Retired
(0.01)***
no
yes
Additional controls
N
4213
4213
R2
0.69
0.74
Results from OLS regressions. The unit of observation is a newspaper by year. All regressions include
year-fixed effects. Robust standard errors, clustered by newspaper, in parentheses. * significant at 10%; **
significant at 5%; *** significant at 1%.
Table 3. NES data summary statistics
Variable
Obs
15014
Congruence
14139
NameRecall
11734
NameRecallIncumbent
11734
NameRecallChallenger
9624
NameRecognition
12459
FeelingThermometerProvided
7441
IdeologicalRatingProvided
10775
LikesOrDislikesProvided
5337
NameRecallSenator
6441
FeelingThermometerSenator
14153
KnowsHouseMajority
14146
KnowsSenateMajority
8985
ReadAboutIncumbent
8957
ReadAboutChallenger
8985
Hear about incumbent on TV
8985
Hear about incumbent on radio
Mean Std. Dev.
0.46
0.29
0.31
0.46
0.29
0.45
0.11
0.32
0.92
0.28
0.82
0.38
0.70
0.46
0.53
0.50
0.46
0.50
0.91
0.29
0.56
0.50
0.50
0.50
0.50
0.50
0.17
0.38
0.50
0.50
0.26
0.44
Min
0.01
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Max
0.97
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
Table 4. Voter knowledge of House Representative
Controls
Fixed effects
Baseline
I
II
no
yes
year
state*year
Dependent variable: NameRecall
0.28
0.35
Congruence
(0.07)*** (0.07)***
Observations
14139
14139
R-squared
0.16
0.27
Within-race
III
IV
no
yes
district*year district*year
Redistricting
V
VI
no
Yes
year,
state*year,
county
county
0.33
(0.08)***
14139
0.21
0.42
(0.07)***
14139
0.30
0.28
(0.07)***
14139
0.10
0.27
(0.06)***
14139
0.24
0.04
(0.06)
9624
0.41
0.10
(0.06)*
9624
0.42
0.15
(0.06)**
9624
0.08
0.07
(0.06)
9624
0.31
Dependent variable: FeelingThermometerProvided
0.20
0.19
0.19
Congruence
(0.05)*** (0.06)***
(0.07)***
Observations
12459
12459
12459
R-squared
0.18
0.25
0.23
0.19
(0.07)***
12459
0.28
0.35
(0.09)***
12459
0.10
0.29
(0.09)***
12459
0.19
Dependent variable: IdeologicalRatingProvided
0.23
0.25
0.25
Congruence
(0.09)*** (0.09)***
(0.10)**
Observations
7441
7441
7441
R-squared
0.18
0.25
0.21
0.30
(0.10)***
7441
0.27
0.32
(0.09)***
7441
0.08
0.20
(0.09)**
7441
0.19
Dependent variable: NameRecognition
0.04
0.08
Congruence
(0.05)
(0.05)
Observations
9624
9624
R-squared
0.27
0.39
Dependent variable: LikesOrDislikesProvided
0.28
0.26
0.24
0.21
0.30
0.30
Congruence
(0.08)*** (0.09)***
(0.10)**
(0.09)**
(0.07)*** (0.07)***
Observations
10775
10775
10775
10775
10775
10775
R-squared
0.17
0.29
0.22
0.32
0.1
0.24
Results from OLS regressions. Robust standard errors, clustered by county, in parentheses: * significant at
10%; ** significant at 5%; *** significant at 1%.
Table 5. Auxiliary regressions. Dependent variable: Congruence
I
II
III
Baseline
Within-race
Redistricting
Explanatory variables:
2(17), 0.13
1(17), 0.22
0(17), 0.24
Individual controls
County controls
5(8), 0.01
4(8), 0.08
0(8), 0.34
Urbanism controls
1(4), 0.08
2(4), 0.02
1(4), 0.08
Race controls
0(3), 0.60
1(3), 0.04
Rep. controls
0(8), 0.18
0(8), 0.57
Fixed effects
State*year,
incumbent
District*year
State*year,
county
Observations
14139
14139
14139
Results from OLS regressions. Each entry of format x(y), z, where x is the number of variables significant at
the 5% level, y is the total number of variables in that category, and z is the p-value of a test of joint
significant of all y variables. Robust standard errors, clustered by county, were used.
Table 6. Placebo: General political knowledge
Dependent NameRecall Feel.Therm. KnowsHouse KnowsSenate
variable
Senator
Senator
Majority
Majority
-0.02
0.04
0.00
0.02
Congruence
(0.09)
(0.14)
(0.05)
(0.06)
6441
Observations
5337
14153
14146
0.22
R-squared
0.30
0.31
0.27
Results from OLS regressions. All regressions include state*year- and incumbent-fixed effects, and controls
(“baseline specification”). Robust standard errors, clustered by county, in parentheses. * significant at 10%;
** significant at 5%; *** significant at 1%.
Table 7. TV and radio
Read/hear about Rep. from
Name
Recall
Newspaper Radio
TV
0.45
0.10
0.19
0.33
(0.10)***
(0.07)
(0.11)*
(0.09)***
-0.29
0.62
0.68
0.17
TVCongruence
(0.22)
(0.19)*** (0.22)***
(0.19)
Observations
8985
8985
8985
14139
R-squared
0.22
0.16
0.24
0.27
Results from OLS regressions. All regressions include state*year- and incumbent-fixed effects, and controls
(“baseline specification”). Robust standard errors, clustered by county, in parentheses. * significant at 10%;
** significant at 5%; *** significant at 1%.
Dependent
variable
Congruence
Table 8. Representatives in Congress summary statistics
Variable
Obs
Mean Std. Dev. Min Max
4768
0.45
0.24 0.00 0.99
Congruence
3317
0.01
0.36 -0.72 0.89
NominateScores
4508
0.27
0.44
0
1
DistribComm
4768
0.16
0.36
0
1
PolicyComm
4553
3.79
3.72
0
28
Witness
4553
1.67
2.20
0
21
WitnessBudget
Table 9. Dependent variable: Number of witness appearances before congressional hearings
I
II
III
IV
V
VI
VII
Congruence
0.45
0.47
0.46
0.45
0.43
0.45
0.43
(0.16)*** (0.15)*** (0.15)*** (0.21)** (0.15)*** (0.22)**
(0.21)**
District controls
x
x
x
x
x
x
x
Race and Representative controls
x
x
x
x
x
Fixed-effects
state,year state,year state,year district,year state*year state,year state,year
Estimation procedure
Poisson
NB
NB
NB
NB
Poisson
NB
Appearance before committee
All
All
All
All
All
Appr. W&M Appr. W&M
Observations
4553
4553
4553
4553
4553
4553
4553
Estimation procedure: NB = negative binomial regression; Poisson=Poisson regression. The unit of
observation is House Representative by congressional session. Appearance before committee:
Appr.=Appropriations; W&M=Ways and Means. Standard errors, clustered by congressional district in
parenthesis: * significant at 10%; ** significant at 5%; *** significant at 1%.
Table 10. Committee assignments
Dependent variable
Distributive committee assignment
Policy committee assignment
I
II
III
IV
V
VI
VII
VIII
0.41
0.11
0.15
0.03
-0.18
-0.10
-0.38
-0.24
Congruence
(0.07)***
(0.09)
(0.09)*
(0.08)
(0.06)***
(0.08)
(0.09)***
(0.08)***
Controls
no
yes
no
yes
no
yes
no
yes
Fixed effects
state*year state*year year, district year, district state*year state*year year, district year, district
Observations
4508
4508
4508
4508
4768
4768
4768
4768
R-squared
0.17
0.36
0.45
0.56
0.12
0.22
0.48
0.53
Results from OLS regressions. The unit of observation is House Representative by congressional session.
Standard errors clustered by House representative: * significant at 10%; ** significant at 5%; *** significant at
1%.
Table 11. Dependent variable: NOMINATE scores first dimension
IV
I
II
III
V
0.17
0.18
0.17
0.14
(Democrats win)*Congruence 0.19
(0.04)*** (0.04)***
(0.05)***
(0.04)***
(0.05)***
-0.68
-0.67
-0.79
-0.67
-0.80
Democrats win
(0.03)*** (0.03)***
(0.08)***
(0.03)***
(0.08)***
-0.04
-0.10
-0.09
-0.05
-0.02
Congruence
(0.03) (0.04)***
(0.04)**
(0.03)
(0.03)
main, urb.
main, urb.
Controls
no
main
interacted
no
interacted
year, district year, district
Fixed effects
no
no
no
Observations
3317
3317
3317
3317
3317
R-squared
0.88
0.90
0.90
0.96
0.96
Results from OLS regressions. The unit of observation is House Representative by congressional session.
All specifications control for 5th order polynomials in Democratic vote share in the House and presidential
elections. The “main” controls denote the direct effect of all our control variables. The “urb. interacted”
controls denote interaction between the Democrats win dummy variable and share urban, density, and the
10 dummy variables for urban and density. Standard errors, clustered by congressional district in
parenthesis: * significant at 10%; ** significant at 5%; *** significant at 1%.
Table 12: Distribution of federal funds across counties, 1984-2004
Dependent variable, log spending per capita
Baseline
Within-race
I
II
III
IV
0.092
0.137
0.064
0.094
Congruence
(0.030)*** (0.027)***
(0.030)**
(0.030)***
Controls
no
Within neighbor-counties
V
VI
0.106
0.096
(0.039)***
(0.038)**
Redistricting
VII
VIII
0.051
0.035
(0.021)**
(0.020)*
yes
no
yes
no
yes
no
district*
district*
year,
year,
state*year,
year
year
neighbor
neighbor
county
Fixed effects
state*year state*year
Observations
33085
33085
28787
28787
16698
16698
33085
R-squared
0.259
0.393
0.441
0.516
0.638
0.677
0.882
Results from OLS regressions. The unit of observation is county by congressional session. Standard errors,
clustered by county, in parentheses: * significant at 10%; ** significant at 5%; *** significant at 1%
Table 13 Voting data summary statistics.
Variable
Obs
17221
Democratic vote share
17221
Incumbent vote share
17221
Congruence
House election turnout
Presidential election turnout
House election turnout (top of ticket)
7106
7106
999
Mean Std. Dev
48
19
66
12
0.49
0.30
40.5
41.6
31
9.3
8.8
9
Min
0
0
0
Max
100
100
1
3.0
4.1
5
99.9
99.5
71
Table 14 Dependent variable:
Percent of population voting in House election, 1984, 1988, 2000
Baseline
Within-race
I
II
III
IV
0.83
0.73
0.83
0.81
Congruence
(0.14)*** (0.14)***
(0.14)***
(0.14)***
0.92
0.89
0.94
0.92
Presidential elec. turnout
(0.01)*** (0.01)***
(0.01)***
(0.01)***
Controls
no
yes
no
yes
Redistricting
V
VI
0.44
0.40
(0.61)
(0.60)
0.84
0.83
(0.03)*** (0.03)***
no
yes
state*year, state*year,
Fixed effects
state*year state*year district*year district*year county
county
Observations
7106
7106
7106
7106
7106
7106
R-squared
0.92
0.92
0.94
0.94
0.95
0.95
Results from OLS regressions. The unit of observation is county by district by election. Standard errors,
clustered by county in parentheses: * significant at 10%; ** significant at 5%; *** significant at 1%.
yes
state*year
, county
33085
0.890
Table 15 Incumbent and challenger effects.
Read about
Name recall
Incumbent Challenger
Incumbent
Challenger
Congruence
0.38
0.05
0.33
0.12
(0.12)***
(0.10)
(0.09)***
(0.05)**
Observations
5945
5930
10424
10424
R-squared
0.24
0.29
0.29
0.30
Results from OLS regressions. Sub-sample of contested elections. All regressions include state*year- and
incumbent-fixed effects, and controls (“baseline specification”). Robust standard errors, clustered by county,
in parentheses: * significant at 10%; ** significant at 5%; *** significant at 1%.
Table 16 Incumbency advantage:
Dependent variable: Democratic percent of two-party vote
Levitt and Wolfram
Baseline
V
VI
9.16
9.71
9.33
8.27
Incumbent
(0.23)***
(0.26)***
(1.25)*** (1.38)***
0.68
0.68
0.74
1.13
Incumbent*Congruence
(0.27)**
(0.26)***
(0.29)** (0.30)***
0.86
0.82
1.50
1.74
Congruence
(0.28)***
(0.27)***
(0.30)*** (0.31)***
0.74
0.74
0.67
0.66
Presidential vote share
(0.01)***
(0.01)***
(0.01)*** (0.01)***
1.22
2.58
Democratic Incumbent
(2.48)
(2.75)
20.72
23.39
Lagged Dem vote
(1.13)*** (1.11)***
Within-race
VII
VIII
0.64
(0.26)**
0.60
(0.27)**
0.76
(0.01)***
0.77
(0.26)***
0.63
(0.26)**
0.75
(0.01)***
Redistricting
IX
X
10.82
12.77
(1.93)*** (1.78)***
1.43
1.21
(0.46)*** (0.44)***
2.57
2.55
(0.74)*** (0.73)***
0.6
0.59
(0.02)*** (0.02)***
-2.57
-5.5
(3.84)
(3.54)
19.51
17.47
(1.42)*** (1.41)***
no
yes
no
yes
no
yes
no
yes
state*year, state*year,
state*year, state*year,
Fixed effects
district*plan district*plan state*year state*year district*year district*year
county
county
Observations
17221
17221
14147
14147
14147
14147
14147
14147
R-squared
0.89
0.9
0.87
0.86
0.94
0.95
0.91
0.92
Results from OLS regressions. The unit of observation is county by district by election. Standard errors,
clustered by county in parentheses: * significant at 10%; ** significant at 5%; *** significant at 1%.
Controls
Congressional district lines
Newspaper A
Newspaper B
0
.2
0
.6
Congruence
.4
.2
1
7.8
7.7
7.6
.6
1
.4
.6
2.4 Politicians
2.2 Newspapers
.2
.8
Reader share (living in district)
0
.2
.6
Congruence
.4
.8
1
(e) Constituency committee work
0
(b) Newspaper coverage
0
0
.2
.2
.6
Congruence
.4
.8
1
.6
Congruence
.4
.8
(c) Voter information
(d) Witness appearances
2.3 Voters
1
Figure 1: Structure of empirical investigation
.4
Congruence
.8
2.5 Policy
(f) Party loyalty
.8
(g) Federal spending per capita
(Circulation-weighted average reader share)
2.1 Congruence
7.9
Log(per capita spending)
7.5
.8
.6
.4
.2
0
Low
95
90
85
80
75
.6
.4
.2
High
Pct. votes with party leaders
1
District share of coverage
Share on constituency comm.'s
0
7
6
5
4
3
.6
.5
.4
.3
.2
(a) Congruence
Appearances per congress
2
Name recall (share correct)
.1
1
Congressional districts
Congruence between newspaper markets
and congressional districts
Congruence
under 0.03
0.03 to 0.08
0.08 to 0.18
0.18 to 0.47
0.47 to 0.73
0.73 to 0.82
0.82 or more
Figure 2: Congruence in Missouri
Congressional districts and
TV markets (DMA’s)
Congruence between TV markets (DMA’s)
and congressional districts
Congruence
under 0.13
0.13 to 0.17
0.17 to 0.20
0.20 to 0.27
0.27 to 0.33
0.33 to 0.48
0.48 or more
Figure 3: TVCongruence in Missouri
1
NominateScores
0
.5
-.5
.4
High Congruence
Fitted values
Figure 4. Nominate Scores by Congruence
Low Congruence
Fitted values
.45
.5
.55
Democratic share of two-party vote in House election
Nominate Scores
.6